Complex sampling does not "break" Nyquist. IQ quadrature sampling produces twice as many bits per second of information (at the same sample rate for real or complex samples), and the 90 degree phase offset between the I and Q channel in those bits provides extra information about the spectrum.
One typical example to demonstrate aliasing is that the samples of DC and the samples of a sinusoid of a frequency at the sample rate look the same. A single channel of samples at 40 MHz would alias 0 Hz and a 40 Hz sinusoid together. Only if you strictly bandlimit the input spectrum to below Fs/2 in bandwidth would you prevent this aliasing. If you feed a strictly real data vector to an DFT, the result would be conjugate symmetric, half the FFT result would be redundant, and the upper half can be called the negative frequency image.
But with quadrature sampling (using a 40 MHz sampling clock plus a 90 degree phased shifted copy), a DC signal and a 40 MHz sinusoid would be clearly differentiated. I and Q would be the same for DC, but different for 40 MHz, e.g. if I is the same, the Q would be different, or vice versa. Therefore the spectrum between half the IQ sample rate and the sample rate will not be aliased with spectrum below, as is the case with single channel sampling. If you feed a DFT using the IQ data as a complex input vector, the lower and upper half of the result can be non-symmetric, thus represent independant information, thus twice the bandwidth (between 0 and Fs), compared to a strictly real DFT of data at the same sample rate.